Object Detection Accuracy via Location Probability Correction

The patent applies local quality by differentiating the treatment of detected objects based on their location probability. High probability locations receive different confidence thresholds or weighting compared to low probability locations. This means the detection system adapts its evaluation criteria locally according to the spatial distribution patterns learned from training data, thereby improving accuracy without uniformly complicating the entire detection process.

The patent changes the confidence threshold parameter dynamically based on the detected object's location probability. Instead of using a fixed threshold for all detections, the system adjusts the threshold according to the probability distribution of object locations for each class. This parameter adaptation resolves the contradiction by making the detection process more accurate while maintaining a relatively simple underlying algorithm structure.

The patent applies local quality by applying different correction factors to confidence scores based on the detected object's location probability. Objects detected in high probability locations receive smaller corrections (or even boosts), while objects in low probability locations receive larger corrections. This localized differentiation allows the system to maintain high precision for reliable detections while preserving recall for potentially missed objects, resolving the trade-off between precision and recall.

The patent introduces dynamics by making the correction factor adaptive rather than static. The correction factor is determined dynamically based on the intersection between the detected bounding box and the probability distribution map for that object class. This dynamic adjustment allows the system to optimize precision-recall balance by adapting the correction strength to each specific detection case rather than applying a uniform correction.

The patent applies preliminary action by pre-computing the probability distribution maps for each object class during the training phase. These maps capture the spatial tendencies of where objects of each class are likely to appear. During inference, the system simply queries these pre-computed maps rather than performing complex real-time analysis, thereby improving accuracy without significantly increasing the complexity of the detection process.

The patent introduces an intermediary component - the probability distribution map - that mediates between the raw detection results and the final confidence evaluation. This intermediary layer translates the spatial location information into probability scores, which then guide the confidence threshold application. This intermediary simplifies the overall process by decoupling the detection logic from the location-aware evaluation logic.